Can Single-Mode Fiber Support the New Foundation of AI Data Centers

Preface: A Connection Revolution that is taking place

If you have been following the news in the telecommunications industry over the past few weeks, you have probably noticed two things: Meta promised Corning 6 billion US dollars to purchase optical fibers and cables for AI data centers. The spot price of G.652.D single-mode optical fiber has soared from less than 20 yuan per core kilometer to over 83 yuan in more than a year. Outside the high-frequency trading stock market, few people realize that we are at a critical juncture in the evolution of network infrastructure architecture - AI is fundamentally changing the way connections are made within data centers.

The global market size of AI-specific optical transceiver modules is expected to soar from 16.5 billion US dollars in 2025 to 26 billion US dollars in 2026, with an annual growth rate of over 57%. This figure reflects the accelerating expansion wave of global AI data centers. In this wave, single-mode optical fibers are penetrating into the interior of data centers in an unprecedented manner, transforming from a "supporting role" in long-distance transmission to the "leading role" in supporting the high-speed interconnection of AI clusters.

Computing power is surging, while bandwidth is catching up from behind

To understand why single-mode optical fibers are worth this much, it is first necessary to understand the network bandwidth predicament faced by AI clusters.

From GPT-3 to GPT-5, the scale of model parameters has jumped from the hundreds of billions to the trillions, and the demand for training computing power has soared from the kilocalories level to the hundreds of thousands of kilocalories level. More importantly, with the emergence of large models of the thinking chain, the computing power demand in the reasoning stage has even reached over a hundred times that of traditional models.

However, the computing power of a single node has increased by approximately 1,000 times in eight years, while the network bandwidth has only risen from 200G to 800G in the past four years. This imbalance of "computing power soaring like a rocket and network advancing like walking" makes network bandwidth the core bottleneck restricting the efficiency of clusters with tens of thousands or even hundreds of thousands of cards during collaborative training.

Thus, a fundamental issue emerged: the connection architecture of the data center needed to be redesigned.

Single-mode and multi-mode: It's not just a difference of "a bit thicker and a bit thinner"

The debate over single-mode optical fibers and multimode optical fibers is actually not new. The core difference between the two lies in the core diameter and transmission mode: The core diameter of single-mode optical fiber is only 8-10 micrometers, which is much smaller than that of multimode optical fiber, which is 50 or 62.5 micrometers. It is precisely this micron-level difference that brings about a fundamental physical advantage - single-mode fibers eliminate modal dispersion fundamentally by confining light to a single fundamental mode for propagation.

What does this mean? This means that there is theoretically no upper limit to the bandwidth-distance product of single-mode optical fibers. In actual deployment, when OS2 single-mode optical fibers are combined with high-speed optical modules such as LR4, DR4, and ZR, the transmission distance can easily exceed 10 kilometers. The effective transmission distance of multimode optical fibers at a rate of 400G is compressed to approximately 100 meters.

In the traditional context of data center design, these physical characteristics determine the basic division of labor of "single-mode for long distances and multi-mode for short distances". But this division of labor is being redefined in the face of AI clusters. The distance between GPU cabinets in an AI training cluster may only be tens to hundreds of meters. However, it is precisely this "middle ground" that makes the transmission distance of multimode optical fibers stretched thin, while single-mode optical fibers, with their flexible power budget, become a more reliable architectural choice.

Deeper Logic: The "Unification" and "Simplification" of Architecture

If the transmission distance is regarded as the first ticket for single-mode optical fibers to enter data centers, then its architectural advantages are the real reasons that attract network architects.

On the one hand, single-mode optical fibers do not require complex parallel structures. When multimode fibers support rates of 400G and above, they usually rely on multi-core parallel solutions (such as SR8, where 8 cores of optical fibers transmit simultaneously), which leads to a significant increase in the number of optical cables per link, thereby triggering a chain reaction of growth in cabling pressure, management complexity, and deployment costs. Single-mode optical fibers can simultaneously transmit multiple wavelength channels in a single fiber through duplex interfaces or wavelength division multiplexing (WDM), achieving rates of 400G, 800G, and even the future 1.6T. This simplicity holds irreplaceable value in high-density AI training networks.

On the other hand, from the perspective of network topology design, the short-distance limitations of multimode fibers will force architects to shorten links, add intermediate switching layers or introduce active relay equipment - these makeshift measures may still be tolerable when there are 1,000 connections, but when the number of connections expands to hundreds of thousands, each small compromise will accumulate into huge architectural complexity and operation and maintenance costs. The long-distance transmission capability provided by single-mode optical fibers precisely offers an "uncompromising" architectural foundation for the dual-dimensional expansion of Scale Out and Scale Up in large computing power clusters.

CPO and Silicon Photonics: The Next Accelerator for Single-Mode Fibers

It is worth noting that the popularization of single-mode optical fibers is not an isolated technological trend. It is complementing and resonating with the breakthroughs in silicon photonics and co-packaged optics (CPO).

In early 2026, NVIDIA released the Spectrum-X silicon photonics CPO switch. Compared with the traditional pluggable optical module solution, CPO directly packages the optical engine with the switching chip, reducing the power consumption of the 1.6Tb/s port to one fifth, improving signal integrity by more than 60 times, and enhancing network reliability by 10 times. Spectrum-X Photonics offers two configurations: 128 800Gb/s ports (with a total bandwidth of 100Tb/s) and 512 800Gb/s ports (with a total bandwidth of 400Tb/s).

From the perspective of the evolution law of connection technology, when the rate climbs to a certain extent, "light" is bound to replace "electricity" as the mainstream interconnection medium between cabinets and even within cabinets. This is precisely one of the most crucial judgments in NVIDIA's platform roadmap from Blackwell to Rubin: The effective transmission distance of copper cables at a rate of 800G can no longer meet the requirements of cross-cabinet GPU collaboration, while the combination of CPO and single-mode fiber is redefining the limits of physical layer connections.

Meanwhile, the IEEE 802.3dj standard is in the D2.4 draft stage and is formulating the standard framework for 800G and 1.6TbE based on 200G/lane signaling. The maturity of the standard will further reduce the cost of single-mode fiber optic modules and accelerate their deployment within data centers.

Future Outlook: From "Substitution" to "Reshaping"

The rise of single-mode optical fibers in AI data centers is essentially a microcosm of the transformation we are undergoing from an "electronic-dominated" to an "optical-electrical integration" network architecture. It resolves the physical bottleneck of data transmission and also makes it possible to build a larger-scale, flatter and more efficient AI cluster architecture.

Meta paid $6 billion to secure Corning's optical fiber production capacity, NVIDIA fully embraced CPO+ silicon photonics technology, and the tender volume of G.654.E optical fiber increased by 25 times in a year - these signals combined outline a clear industry trend line: The optical interconnection architecture of AI data centers is accelerating its migration from "multi-mode in-main and single-mode out-of-main" to "single-mode full coverage".

For architects planning the next-generation AI data centers, the question is no longer "whether to deploy single-mode fiber internally", but "when to start deploying". From 2025 to 2026, the price of single-mode optical fibers has soared by more than four times in just over a year. This is not due to speculative speculation but rather a structural squeeze on infrastructure supply caused by the demand for AI computing power. In this wave of optical infrastructure construction driven by AI, whoever can complete the architectural upgrade from multi-mode to single-mode first will gain a first-mover advantage in the competition of AI with larger scale, lower latency and higher energy efficiency.

From a longer time perspective, this is not only an evolution of optical fiber types, but also a shift in the design paradigm of the physical layer in data centers. In the past, using multiple modes for short distances was "common sense". In the face of the GPU collaboration in AI training clusters often in the tens of thousands, the increasingly expanding model parameters, and the continuously rising data round-trip frequency, single-mode fiber provides exactly what the current "hunger" for computing power urgently needs - nearly unlimited scalability, extremely low operation and maintenance complexity, and the "headroom" for the future 1.6T and even 3.2T era. It is precisely because of this "design once, worry-free for ten thousand cards" architectural capability that optical fibers and cables are transforming from insignificant "accessories" in traditional data centers to the most indispensable strategic resources in AI clusters.